From CpG-Level Effects to Regions
A methylation array measures methylation at individual CpG sites, and differential testing estimates a group difference at each site. A regional analysis asks a different question: do multiple nearby CpGs, considered together, show a pattern of methylation difference between groups? The answer depends partly on the data and partly on how the analyst defines a region. A set of significant CpGs is not automatically a uniquely defined genomic feature.
Suppose the site-level analysis reports, for each CpG \(j\), its genomic coordinate, an estimated group difference, and a measure of uncertainty or evidence. For clarity, let \(d_j\) denote the adjusted difference in mean beta value at CpG \(j\), calculated as the mean for cases minus the mean for controls. Thus \(d_j=-0.06\) represents 6 percentage points lower methylation in cases at that site. The covariate adjustment and comparison must be defined consistently across the sites being combined.
Before forming regions, the site data need to be suitable for the intended comparison. CpG measurements should have passed probe and sample quality checks, coordinates should refer to the same genome build, and the site-level effects should represent the same group contrast with compatible covariate adjustment. A region rule should also specify how many CpGs are required, how far apart they may be, and whether intervening sites without evidence of change can be included. These choices can change both the reported boundaries and the number of regions.
Common Ways to Define a Region
Three broad approaches are useful to distinguish. They answer related, but not identical, questions. In each case, the rule should be described clearly enough that another analyst could reproduce which CpGs belong to each region.
| Approach | How regions are formed or tested | What the result emphasizes | Main caution |
|---|---|---|---|
| Threshold-and-merge | Select CpGs using a stated site-level evidence threshold, then group selected sites within a specified maximum gap. | Clusters of individually supported sites. | Boundaries and results can be sensitive to the threshold and gap rule; a selected set is not itself a calibrated region-level test. |
| Fixed windows or tiles | Partition coordinates into intervals of prespecified width, or test windows that move along the genome. | Differences within a coordinate-defined interval. | Results can depend on window width, starting position, and how sites near boundaries are handled. |
| Region-level testing | Define candidate intervals and test a summary or joint signal across their CpGs using a method designed for regional inference. | Evidence for a regional contrast, which need not require every CpG to pass a site-level threshold. | The test needs a valid uncertainty calculation and adjustment for the number of regions considered. |
A related strategy smooths site-level estimates or segments the genome according to changes in their pattern. It can make a gradual pattern easier to summarize, but the smoothing or segmentation rule also affects boundaries and uncertainty. Regardless of the construction method, report the criteria used rather than describing a region as though it had been discovered without analytical choices.
Worked Example: How a Gap Rule Changes a Merged Region
Consider hypothetical CpGs at positions 1,000, 1,045, 1,190, 1,280, and 1,405 base pairs within one genomic span. Their site-level p-values are 0.002, 0.008, 0.20, 0.004, and 0.009, respectively. Suppose the rule selects sites with \(p<0.01\), joins selected sites when the distance between consecutive selected coordinates is at most a chosen gap, and requires at least three selected CpGs per reported region.
The selected coordinates are 1,000, 1,045, 1,280, and 1,405. With a maximum gap of 150 base pairs, the distances between consecutive selected sites are \(1{,}045-1{,}000=45\), \(1{,}280-1{,}045=235\), and \(1{,}405-1{,}280=125\) base pairs. This forms two clusters: the first contains two selected CpGs and the second contains two. Neither meets the three-CpG minimum, so this rule reports no DMR.
Now increase the maximum gap to 250 base pairs. Each of the three distances is at most 250, so all four selected CpGs form one cluster. The rule reports a region from position 1,000 to position 1,405, meeting the minimum of three selected sites. The CpG at 1,190, whose p-value is 0.20, lies inside those boundaries but was not one of the selected sites. This example shows why a report should state whether a region spans intervening unselected CpGs and how the maximum gap was applied.
Defining and Interpreting a Regional Effect
A region-level effect needs an estimand: the precise quantity the estimate is intended to summarize. One transparent choice is the equal-weight average of the site-level differences across the \(m\) CpGs assigned to a region \(R\). If \(d_j\) is the adjusted case-minus-control difference in beta value at site \(j\), define:
Because beta values are proportions from 0 to 1, this regional estimate is also on the beta-value scale. Multiplying by 100 expresses it in percentage points. Equal weighting gives each included CpG the same contribution; it does not mean that each site has the same measurement precision or biological importance. Other definitions use prespecified weights, a joint model, or a summary on the M-value scale. These are different estimands and should not be described as interchangeable.
The averaging formula follows directly from defining a regional contrast as the difference between the average case and average control methylation across the included sites. Let \(\hat{\mu}_{1j}\) and \(\hat{\mu}_{0j}\) be the adjusted group means at site \(j\), so \(d_j=\hat{\mu}_{1j}-\hat{\mu}_{0j}\). Then:
This identity justifies the estimate when the same sites and equal weights are used for both groups. It does not, by itself, provide a standard error or a test. In particular, a regional standard error must reflect uncertainty in the site estimates and their joint behavior. A valid region-level method supplies an uncertainty calculation appropriate to its model; averaging site-level standard errors is not a substitute.
Worked Example: Calculating an Equal-Weight Regional Difference
For a hypothetical region containing four CpGs, suppose the adjusted case-minus-control differences in beta value are \(-0.12\), \(-0.08\), \(-0.04\), and \(+0.04\). Calculate the equal-weight regional estimate.
There are \(m=4\) included sites, so:
The estimated regional average is \(-0.05\) beta units, or \(-5\) percentage points. Under this definition, the average methylation difference across these four sites is estimated to be 5 percentage points lower in cases than controls. The positive difference at the fourth site shows that the region does not have a uniformly negative site-level pattern. The average summarizes the chosen sites; it does not imply that every CpG is lower in cases or that every site has a statistically supported difference.
Regional Evidence and Its Assumptions
A regional effect estimate and a test of regional evidence are separate outputs. The estimate describes the size and direction of the defined contrast. A test or interval describes uncertainty under a stated model. For a model-based regional test, the analysis needs a valid null distribution for the regional statistic, appropriate handling of covariates, and adjustment for the number of regions tested. If regions are selected using the same data, a method that accounts for that selection or otherwise provides calibrated inference is needed; an ordinary p-value calculated after selection may overstate evidence.
The threshold-and-merge approach is particularly easy to misinterpret. A merged cluster of sites that each passed a site-level threshold may be useful for locating a pattern, but it does not automatically have a valid region-level p-value. Conversely, a region-level test may identify a collective pattern even when some individual CpGs do not cross a site-level threshold. Always identify whether the reported claim concerns a collection rule, a regional test, or both.
Worked Example: Interpreting a Region-Level Estimate and Test
Suppose a hypothetical fixed-window analysis defines a region in advance and reports an adjusted regional difference of \(-0.065\) beta units, with a model-based standard error of \(0.020\). Assume the regional method’s stated model is appropriate and its standard error is calibrated for the observations and sites included. For illustration, use a large-sample normal interval and two-sided test of \(H_0:\Delta_R=0\).
The 95% interval is estimated as the regional difference plus or minus 1.96 standard errors:
The interval is approximately \(-10.42\) to \(-2.58\) percentage points on the beta-value scale. Under the assumed model, it is a 95% confidence interval for the defined regional contrast, not a 95% probability statement about this particular interval containing the parameter.
The test statistic is \(z=-0.065/0.020=-3.25\), giving an approximate two-sided p-value of 0.00115. This is evidence against a zero regional contrast under the stated test model. It does not establish a causal effect, prove that every CpG in the window changes, or show that the interval is significant after testing many regions. The final inferential claim must use the analysis’s multiplicity adjustment across the regions considered.
Diagnostics, Comparisons, and Limitations
Useful checks begin with the region definition and the underlying CpGs. Confirm that the coordinates use a consistent genome build, that each reported region contains the required number of valid sites, and that the listed boundaries follow the declared gap or window rule. Inspect the site-level effects and their directions: a moderate regional average can conceal strong differences in opposite directions, while a coherent-looking pattern may still be uncertain. Check that the regional estimate is on the stated scale and that the group contrast and covariate adjustment match the site-level analysis.
No single construction rule is best in every setting. Threshold-and-merge rules are straightforward to explain and can focus attention on clusters of individually supported sites, but results can change when thresholds or gap limits change. Fixed windows make intervals comparable and reproducible, but a signal near a window boundary can be split or diluted. Region-level tests can directly target a defined regional contrast, but their interpretation relies on the test’s model and multiplicity procedure. Smoothing and segmentation can represent continuous patterns, but introduce additional choices about how much local variation to smooth and where to place boundaries.
A common failure is to treat nearby CpGs as independent evidence simply because they have separate rows in a data table. They are measured at distinct sites, but their estimates need not be statistically independent. A method that combines their evidence must use an uncertainty calculation suitable for the joint data, rather than multiplying p-values or adding standard errors as if dependence were absent. The detailed structure of dependence is a separate issue; here, the practical requirement is not to assume independence without justification.
A further limitation is that region boundaries depend on array coverage: only assayed, usable CpGs can contribute to an array-based region. The observed interval is therefore not necessarily a complete map of methylation across every CpG in the genomic span. Also, an average difference can hide variation among sites, and a significant regional result does not identify a molecular mechanism. Describe the assay, region rule, effect scale, and uncertainty alongside any biological interpretation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- Why must a threshold-and-merge result report both its site-level threshold and its maximum-gap rule?
- Three CpGs have case-minus-control beta differences of \(-0.03\), \(-0.09\), and \(+0.03\). What is their equal-weight regional difference, in beta units and percentage points?
- Does a regional average of \(-0.05\) imply that every CpG in the region has a negative difference? Explain.
- What is the error in calling a merged set of individually significant CpGs a significant region without a region-level inferential procedure?
Solutions. (1) Both choices determine which sites are grouped and therefore can change the boundaries and number of reported regions. (2) The average is \((-0.03-0.09+0.03)/3=-0.09/3=-0.03\) beta units, or \(-3\) percentage points. (3) No. An average can be negative even when one or more included sites have positive differences; it summarizes the included values rather than asserting a common direction at every site. (4) A site-selection and merging rule does not itself define a calibrated region-level test. Regional inference needs a suitable uncertainty calculation and adjustment for the regions considered.